Hermes Agent and OpenClaw both built real software this year. OpenClaw crossed 345,000 GitHub stars in early April and now carries more than 5,700 community-contributed skills. Hermes Agent passed 64,000 stars on a much narrower pitch: an agent that gets better at the work you actually do. Most of the comparison posts already out there miss the question that matters, which is what shape of agent you're building. That's the question this guide is for.
Both are open-source. Both self-host. Neither is a fork of the other; they're separate codebases solving overlapping problems with different philosophies. If you read one of the recent vs-pages on KuCoin or Kanaries and walked away thinking "these tools are interchangeable," they're not. Picking the wrong one for your workload can costs weeks.
The two architectures, in one paragraph each
OpenClaw is gateway-centric. A long-running Gateway process sits on a host (Linux, macOS, or Windows) and routes work between channels (Telegram, Discord, Slack), nodes (your phone, laptop, workstation), and skills. The 5,700+ skill marketplace is its center of gravity. The agent's intelligence comes from composing skills, not from learning new ones. Multi-agent swarming over ACP is built in. The mental model is: assistant control plane.
Hermes Agent is loop-centric. The runtime is built around a closed feedback cycle: the agent attempts a task, scores its own performance, distills the successful path into a reusable skill file, and gets measurably better next time. NousResearch's tagline, "the agent that grows with you," describes how it actually works. The integration surface is smaller than OpenClaw's by design. The mental model is: deep specialist that compounds.
Side-by-side comparison
| Dimension | OpenClaw | Hermes Agent |
|---|---|---|
| Runtime model | Gateway + nodes + channels | Single agent runtime with skill loop |
| Memory architecture | Channel-scoped sessions, conversation logs | Persistent user model, skill files, evolving entity store |
| Skills | 5,700+ community-contributed | Self-generated from successful task completions |
| Default channels | Telegram, Discord, Slack out of the box | CLI; channels via adapters |
| Multi-agent | Yes (ACP swarm) | Sub-agents with namespace isolation |
| Officially supported OS | Linux, macOS, Windows | Linux, macOS, WSL2, Termux (Android) |
| Official iOS support | No (node-only iPhone clients exist) | No |
| License | Open source | Open source |
| GitHub stars (Apr 2026) | 345,000+ | 64,000+ |
| Releases shipped | 82 | 6 |
| Disclosed CVEs (2026 YTD) | 9 in 4 days, March 2026 (one CVSS 9.9) | 0 |
| Exposed public instances | 135,000+ across 82 countries (March 2026) | Not separately tracked |
| Self-host complexity | Higher: Docker, YAML, channel webhooks | Lower: CLI install, env-based config |
| Best at | Breadth of integration, multi-channel ops | Depth of learning on repeated tasks |
Two numbers in that table deserve a second look: CVEs and releases. OpenClaw's 82 releases mean velocity, and they also mean breakage. In March 2026, nine CVEs landed in four days, the worst at CVSS 9.9, and Shadowserver found 135,000+ exposed Gateways across 82 countries in the same window. If you self-host OpenClaw, you patch on a tight cadence or you don't run it on the public internet. Hermes hasn't reported an agent-specific CVE yet, but it also has fewer eyes on it. Smaller surface, fewer disclosures so far.
What the community actually says
Reddit threads carry more signal than vendor blog posts here. A recent synthesis of those threads sorts usage roughly like this: about 35% of operators stick with OpenClaw, about 30% have switched to Hermes, about 20% run both side by side, and about 15% don't trust Hermes yet, mostly because a few users have flagged what looks like coordinated promotion by new accounts.
The top OpenClaw complaints repeat across threads. The most-upvoted one (305 votes) reads: "Every single update ships more bugs and problems than before." Roughly a quarter of updates reportedly break response delivery on at least one channel. The second pain point is memory drift; agents forget prior instructions and repeat mistakes. The third is self-host friction: more time spent on Docker, SSH, and YAML than on the actual workflow.
Hermes flips two of those three. Setup is smoother. One user put it: "I am actually getting stuff done instead of debugging." Default memory works better out of the box. Hermes has its own pattern of complaints though: unreliable self-evaluation (the agent sometimes thinks it succeeded when it didn't), auto-improvements that overwrite manual edits to skill files, and a small integration ecosystem. With only 6 releases.
When OpenClaw is the right pick
Pick OpenClaw if your work is shaped like routing. An agent that listens on three channels, dispatches to specialized sub-agents, runs deterministic cron jobs, and pulls from an existing community skill for nearly any third-party API you can name; OpenClaw gets you there faster than anything else. 5,700 skills means most of what you'd want already exists. The Gateway design also gives you stronger control over what the agent is allowed to do, which matters when stakeholders ask.
Common fits: customer-support routing across Telegram and Slack, on-call automation, multi-tenant agent fleets, anything where "channels and integrations" describes more of the surface than "deep task expertise."
When Hermes is the right pick
Pick Hermes if your work is shaped like repetition. If the agent does the same class of task over and over (code review on a specific repo, research synthesis on a defined topic, content QA against your own style guide), Hermes's skill loop pays off in a way OpenClaw's static composition cannot. Benchmarks reported in early-2026 evaluations show agents using self-generated skills completing complex research and code-execution tasks roughly 40% faster than fresh, non-learning instances.
The security posture is also tighter. Container hardening, namespace isolation for sub-agents, and credential rotation are part of the architecture rather than bolted on. Zero disclosed agent-specific CVEs as of April 2026 isn't a guarantee, but it's a better starting point than OpenClaw's March cluster.
Common fits: a solo developer running a long-lived coding task, research workflows, content operations where the same review happens a hundred times, anything where you'd rather have one agent that compounds than ten that don't.
The "run both" case
About 20% of operators in the community sample run them together. The setup is usually OpenClaw as the channel-and-routing layer up front, with Hermes as the deep-skill specialist behind it. Messages land in Telegram or Slack, OpenClaw routes them to the right sub-agent or skill, and for the workloads where compounding matters, the dispatched call ends up at a Hermes agent that has been getting better at exactly that task for weeks.
You pay the maintenance cost of two runtimes, and you still have to design the handoff. For teams that need both broad integration and deep specialization, that cost beats picking one and accepting the gap.
The honest trade-offs
Neither framework solves portability. Skill formats are framework-specific in both directions; an OpenClaw skill won't run on Hermes and a Hermes skill file won't drop into an OpenClaw Gateway. If you invest a year of skill authorship into one, you're not migrating that work later without rewriting it. Pick the runtime you're prepared to live with.
Update churn cuts the other direction. OpenClaw's 82 releases mean forward motion and frequent breakage; Hermes's 6 releases mean stability, a smaller bug surface, and long waits for fixes. On Hermes, keep a human in the loop on self-evaluation. The agent will tell you it succeeded when it didn't. And neither framework has a credible cross-framework skill standard yet. MCP at the Linux Foundation is the most plausible vehicle for one, but it doesn't exist today.
What to do this week
If you haven't picked one, spend a day with each. Install OpenClaw with one channel and two community skills; install Hermes against a task you've done by hand five times this month. Within an hour you'll know whether you want routing or compounding. No third shape of agent splits the difference cleanly yet.
If you want to deploy Hermes, OpenClaw, or both, and switch between them per task, that's the workflow Onepilot packages on iPhone — see also the Hermes skill triage from iPhone guide and the OpenClaw evergreen setup if you're standing one up this week. For the full three-layer stack including Paperclip, see the orchestrator overview.